PaperReading

FoveaBox: Beyond Anchor-based Object Detector

April 2019

tl;dr: RetinaNet based anchor-free method to predict center and top-left/bottom-right corners. It is more robust in detecting long objects.

Overall impression

This paper is one of the many recent papers about anchor-free object detection. FoveaBox and FSAF both build upon RetinaNet, but FSAF selects feature level by learning, but FoveaBox uses overlapping levels to detect an object. FoveaBox can also be used with anchor-based RetinaNet results, which performs better than FSAF.

Key ideas

• Anchor box can be regarded as a feature-sharing sliding window scheme.
• We human naturally recognize the object in the visual scene without enumerating the candidate boxes. For human eyes, the center of the vision field is with the highest visual acuity.
• In RetinaNet, each cls network predicts KA numbers (A=9, number of anchor boxes), and each reg network predicts 4A numbers at each bin of the feature map. In contrast, FoveaBox predicts K and 4 numbers for each feature map bin. This is very similar to FSAF.
• It keeps the scale assignment to different levels of FPN to stabilize training. This is a key difference to YOLOv1. This is very similar to an anchor-box scheme. In a way, FoveaBox still relies on some priors.
  • Cls: A donut shaped exclusion area with shrunk factors of 0.3 and 0.4 of the diameter is used to separate the positive samples and negative samples.
  • Reg: still needs to scale offset to around 1 for faster and more stabilized training.
• Ablation:
  • Increasing beyond 6-9 anchors does not leads to improvement.
  • Having overlap in scale assignment helps. Two adjacent levels are responsible for the same object.
  • Anchor free method is more robust as it is not adapted to the anchor box statistics.
  • FoveaBox can be used to replace FPN.

Technical details

• We could use SoftNMS and bbox voting for post-processing.
• FoveaBox can be used to replace FPN.
• FoveaBox can be used with anchor-based RetinaNet.
• 42.1 AP with ResNeXt-101, but this is single scale. The authors did not report multi-scale test results.

Notes

• Another concise description of anchor boxes in object detection.

  Anchor method suggests dividing the box space (including position, scale, aspect ratio) into discrete bins and refining the object box in the corresponding bin.